Tsinghua University and other universities launch the first open source large model watermarking toolkit MarkLLM, which supports nearly 10 latest watermarking algorithms

The AIxiv column is a column where this site publishes academic and technical content. In the past few years, the AIxiv column of this site has received more than 2,000 reports, covering top laboratories from major universities and companies around the world, effectively promoting academic exchanges and dissemination. If you have excellent work that you want to share, please feel free to contribute or contact us for reporting. Submission email: liyazhou@jiqizhixin.com; zhaoyunfeng@jiqizhixin.com

##This article is sponsored by Tsinghua University, Shanghai Jiao Tong University, University of Sydney, UCSB, Chinese University of Hong Kong, and Hong Kong University of Science and Technology , Hong Kong University of Science and Technology (Guangzhou) jointly completed. The main authors include: Pan Leyi (first author), an undergraduate student at Tsinghua University, whose research direction is large-scale model watermarking; Liu Aiwei, a doctoral student at Tsinghua University, whose research direction is secure and trustworthy large-scale models; He Zhiwei, a doctoral student at Shanghai Jiao Tong University, research His research direction is large model watermarking, large model intelligence, etc.; Gao Zitian, an undergraduate student at the University of Sydney, research direction is large model watermarking; Zhao Xuandong, UCSB PhD candidate, research direction is trustworthy generative AI, etc.; Hu Xuming, Hong Kong University of Science and Technology/Hong Kong Science and Technology He is an assistant professor at Tsinghua University (Guangzhou), and his research interests include secure and trustworthy large models, information extraction, etc. Wen Lijie is a permanent associate professor at Tsinghua University, and his research interests include process mining and natural language processing.

This article introduces an open source model printing algorithm jointly launched by Tsinghua University and other universities. MarkLLM provides a unified model printing algorithm implementation framework, intuitive printing algorithm mechanism visualization, examples, and systematic evaluation modules, aiming to enable researchers to easily experiment, understand, and evaluate the latest printing technology developments. Through MarkLLM, the author hopes to deepen the public's understanding of model printing technology while providing convenience to researchers, and promote the development and promotion of related research.

• Paper name: MarkLLM: An Open-Source Toolkit for LLM Watermarking

• Paper link ：https://arxiv.org/abs/2405.10051

• Code repository: https://github.com/THU- BPM/MarkLLM

The development status of large model watermarking technology & the problems still faced

Large model watermarking is a recently emerging technology. It is embedded in the process of model generation text. Enter specific characteristics to realize the identification and source tracing of organic text. It can be used in scenarios such as fake news detection, maintaining academic integrity, and data and model copyright protection.

The current mainstream large model watermarking algorithm is to embed watermarks in the large model inference stage. This type of method is mainly divided into two major algorithm families:

• KGW family: add watermark by pre-scoring vector, divide the word list into red and green lists, add bias to green words, so that the output prefers green words;

• Christ family: After the scoring vector is generated, a pseudo-random number is used in the pre-sampling process to make the watermark text more relevant to the random number, thereby embedding the watermark.

However, like all emerging technologies, large language model watermarking technology also faces some challenges in use and understanding.

1. How to easily use various large model watermark algorithms to add and detect watermarks?

Various large model watermarking algorithms continue to emerge. However, their implementation is mostly based on the author's own needs and lacks unified class and calling interface design, which requires researchers and the public to invest a lot of effort in using and reproducing these algorithms.

2. How to intuitively understand the internal mechanism of each large model watermarking algorithm?

The underlying mechanism of the large model watermarking algorithm is relatively complex, involving intervention in the scoring vector generation and sampling process in the text generation process of the large model, which is not easy for researchers and the public to understand.

3. How to conveniently and comprehensively evaluate various large model watermarking algorithms?

The evaluation perspectives and indicators are diverse (including detectability, robustness, impact on text quality, etc.), and one evaluation involves multiple steps, making it extremely challenging to comprehensively and quickly evaluate algorithm performance.

MarkLLM: The first open source large model watermarking multifunctional toolkit

In response to the three problems just mentioned, the author designed and implemented a language-oriented MarkLLM, a tool package for speech model watermarking technology.

The main contributions of MarkLLM can be summarized as follows:

1. Functional perspective

##Unified large model watermarking Algorithm implementation framework: supports 9 specific algorithms of two key watermark algorithm families (KGW family and Christ family).
Consistent, user-friendly top-level calling interface: 1 line of code to implement various operations such as adding watermarks and detecting watermarks.

• Customized large model watermarking algorithm mechanism visualization solution: users can visualize different large models under various configurations The internal mechanism of the model watermarking algorithm.

• ##Comprehensive and systematic large-model watermarking algorithm evaluation module: Contains a total of 12 modules covering 3 evaluation angles An assessment tool and two types of automated assessment pipelines.

2. Design perspective: Modular, loosely coupled architecture design, Extremely scalable and flexible.

3. Experimental perspective: The author uses MarkLLM as a research tool and conducts 3 comprehensive experiments from evaluation perspectives on the 9 supported algorithms to prove the practicality of MarkLLM. While being useful, it also provides valuable data reference for subsequent research.

4. Impact on the open source community: MarkLLM has received a lot of attention since it was launched on GitHub, and currently has 140 + stars, and attracted peers to make code contributions through Pull Request, and to communicate and discuss in the issue column.

The author sincerely hopes that the MarkLLM toolkit will not only provide convenience for researchers, but also improve the public’s understanding and participation in large language model watermarking technology. , promote consensus on this technology between the academic community and the public, promote the further development of research and application of large language model watermarking, and contribute to the safe use of large language models.

The author sincerely welcomes everyone to provide valuable opinions, exchange and learn from each other, and also welcomes code contributions through pull requests, so as to maintain a better large-model watermarking technology ecosystem through everyone's joint efforts!

The above is the detailed content of Tsinghua University and other universities launch the first open source large model watermarking toolkit MarkLLM, which supports nearly 10 latest watermarking algorithms. For more information, please follow other related articles on the PHP Chinese website!